Tool Discovery

When a module has many tools (10+), sending all tool schemas in every prompt wastes context and can confuse the LLM. Tool Discovery solves this by letting the LLM search for tools on demand.

Enabling Tool Discovery

Set tool_discovery = "search" on your Module:

class Agent(Module):
    """Agent with many tools."""
    tools = [Gmail(token="..."), Memory(path="./mem.db"), Slack(token="...")]
    tool_discovery = "search"
    max_steps = 15
    final_output = Output

How It Works

With tool_discovery = "search":

1. The LLM prompt includes only search_tools and __finish__ tool schemas (not all tool schemas)
2. The LLM calls search_tools(query="send email") to find relevant tools
3. Matching tool schemas are returned to the LLM
4. The LLM can then call those tools normally
5. All tools remain callable at any time — only their schemas are deferred

When to Use It

• Many tools (10+): Reduces prompt size and helps the LLM focus
• Multiple services: When combining several services with overlapping functionality
• Dynamic tool sets: When tools are added/removed during execution

How Search Works

The ToolRegistry indexes tools by name and description using keyword matching:

• Tokenizes tool names and descriptions into searchable words
• Scores results by token overlap with the query
• Boosts matches on tool names over description-only matches
• Returns top results sorted by relevance

No external dependencies — pure Python keyword search.

Example

from acorn.services.memory import Memory

class KnowledgeWorker(Module):
    """Agent that manages knowledge across multiple services."""

    tools = [
        Memory(path="./knowledge.db"),
        Gmail(token="..."),
        Slack(token="..."),
        Jira(token="..."),
    ]
    tool_discovery = "search"
    max_steps = 20
    final_output = Report

# The LLM sees only search_tools and __finish__ initially.
# It searches for tools as needed:
#   search_tools(query="save note") -> returns memory__save schema
#   search_tools(query="send message slack") -> returns slack__send schema